01-2013-ps-klif_reduced.pdf (5.316Mb)

(1)

Evaluation of new data sources for improving the estimation of background concentrations in Norway

Philipp Schneider and Andrzej Obracaj

Scientific report

(2)

(3)

List of Figures

1 Matrix visualization of NO₂at stationNO0075A Barnehagen. . . 11 2 Comparison of information content . . . 12 3 Maps of 2007 (NO₂) and 2008 annual means of NO₂, O₃, PM₁₀, and

PM_2.5as they were computed in the 2011 study . . . 13 4 Map showing the 2009 average NO₂ concentration measured at all

Airbase background stations. (from Schneider et al. (2011)) . . . 15 5 Map of background stations used for mapping and temporal decompo-

sition . . . 17 6 Example of an empirical semivariogram ˆγ(h)and its model . . . 23 7 Schematic of the general methodology . . . 24 8 Annual mean NO₂concentration for the year 2009 derived from the

OMNO2e daily 0.25°×0.25° product . . . 26 9 Annual mean NO₂concentration for the year 2009 derived from the

SCIAMACHY/TEMIS monthly 0.25°×0.25° product . . . 27 10 Difference image of the mean annual NO₂ column retrieved from

SCIAMACHY and OMI . . . 28 11 The 0.1°×0.1° resolution OMNO2e product over Europe. Shown here

is the 2009 annual mean tropospheric NO₂concentration . . . 29 12 Comparison of the 0.25°×0.25° resolution OMNO2e product (top)

with the 0.1 degree resolution OMNO2e product (bottom), shown for the Po valley region in Northern Italy . . . 30 13 Annual average tropospheric NO₂column for the year 2009 over Nor-

way. Derived from the 0.1 degree high-resolution OMNO2e product . . 31 14 Scatterplot of Airbase-derived annual mean 2009 station NO₂ con-

centration against the 2009 annual mean tropospheric NO₂columns derived from the OMNO2e high-resolution product. . . 32 15 Empirical and modeled semivariogram of the residuals . . . 33 16 Map showing the residuals from the model fitted between the aver-

age 2009 NO₂ at at all Airbase background stations and the mean tropospheric NO₂column provided by the high-resolution OMNO2e product. . . 34 17 Map showing the residuals from the model fitted between the average

2009 NO₂at at all Airbase background stations in Norway and the mean tropospheric NO₂column provided by the high-resolution OMNO2e product . . . 35 18 Map showing the geostatistically interpolated residuals given at the

station level in Figure 16 . . . 36 19 The average NO₂concentration in Norway for 2009 . . . 37 20 Annual mean NO₂concentration over southern Norway as computed

by the CHIMERE chemical transport model. . . 38 21 Comparison of 2009 hourly time series of NO₂from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 40 22 Comparison of 2009 hourly time series of NO₂from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations, here shown as a two-dimensional histograms (or density plots). . . 41

(6)

23 Comparison of 2009 hourly time series of O₃from station measurements in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 42 24 Comparison of 2009 hourly time series of O₃from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations, here shown as two-dimensional histograms (or density plots). . . 43 25 Comparison of 2009 hourly time series of PM₁₀from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 44 26 Comparison of 2009 hourly time series of PM₁₀from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations, here shown as a two-dimensional histograms (or density plots). . . 45 27 Comparison of 2009 hourly time series of PM_2.5from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 46 28 Comparison of 2009 hourly time series of PM_2.5from station measure-

ments in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations, here shown as two-dimensional histograms (or density plots). . . 46 29 Screenshot of the mapping component of the online web mapping

application . . . 49 30 As Figure 29, here showing background concentrations of PM₁₀in the

greater Oslo area . . . 50 31 As Figure 29, here showing background concentrations of O₃in the

Lillehammer area. . . 51 32 Example screenshot of downloaded data . . . 52

(7)

List of Tables

1 Overview of Norwegian background air quality stations that were used for temporal characterization . . . 16 2 Overview of station type and components measured at each station as

well as their respective long-term mean . . . 18 3 Result of simple linear regression between the 2009 hourly time series

of NO₂from station measurements in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 39 4 Result of simple linear regression between the 2009 hourly time series

of O₃from station measurements in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 42 5 Result of simple linear regression between the 2009 hourly time series

of PM₁₀ from station measurements in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 44 6 Result of simple linear regression between the 2009 hourly time series

of PM_2.5 from station measurements in southern Norway and the corresponding time series extracted from the CHIMERE output over the same locations. . . 46

(8)

(9)

Summary

Based on the experience gained as part of a previous project on developing an “atlas”

of background concentrations for NO₂, O₃, PM₁₀, and PM_2.5over Norway (Schneider et al., 2011), additional work has been carried out in order to evaluate potential improvements to the existing dataset. Both the previous work and this follow-up project were funded by the Climate and Pollution Agency of Norway (KLIF). Three major objectives were addressed as part of this work:

1. To evaluate the potential of satellite-derived NO₂data for improving station- based NO₂mapping in Norway

2. To evaluate the feasibility of using data from a high-resolution atmospheric model for better spatial and temporal characterization of background concentrations

3. To make the results available through an online web mapping system to allow easy access to the data and to provide basic visualization of the results

In order to evaluate the potential of satellite data for mapping air quality, different satellite products providing information on NO₂were first tested and compared. An experimental high-resolution product acquired by the Ozone Mapping Instrument (OMI) on board of the National Aeronautics and Space Administration’s Aura platform was identified as the most suitable product with respect to its similarity to future datasets to be acquired by satellites associated with the Global Monitoring for Envi- ronment and Security (GMES) initiative. This product was then used as an auxiliary dataset to guide the spatial interpolation of NO₂station data. The results indicate that the satellite dataset is useful in providing information on spatial patterns in areas with very low station density such as Norway. Using a simple cross-validation scheme it was shown that a kriging procedure involving OMI-based auxiliary data allowed for mapping NO₂with a lower overall root mean squared error than when using ordinary kriging on station data alone.

In a second task, the output from a high-resolution run of the chemical transport model CHIMERE was evaluated with respect to its ability to contribute to estimating the background concentrations over Norway. This was accomplished using a compar- ative analysis of station observations with CHIMERE-derived time series at the same location. Using a linear regression model the correlation between the two datasets was then studied for several stations and atmospheric pollutants. The results indicate that, with exception of O₃, the CHIMERE-derived hourly time series are generally only weakly correlated with the actual station observations, at least with respect to high-frequency temporal variability. This fact, combined with the lack of long time series of high resolution model output, makes the use of this dataset challenging for temporal characterization of the background concentrations. However, the data is still very valuable as an auxiliary dataset for providing spatial information to assist a geostatistical interpolation of station data.

Finally, in the third task of this study, a web-based mapping portal was developed in order to provide flexible access to the data and to basic visualization tools. The web site provides a Geoserver-based environment for exploring the data by freely zooming and panning in a map interface and further offers the possibility for displaying and downloading customized time series for the various pollutants at any point within Norway.

(10)

(11)

1 Introduction and Background

Knowing the average or typical background concentration of atmospheric pollutants such as O₃, NO₂, PM₁₀, and PM_2.5at a given point in space and time is critical for a variety of applications, and in particular for a comprehensive assessment of air quality.

For this reason, the Climate and Pollution Agency of Norway (KLIF) contracted the Norwegian Institute for Air Research (NILU) in 2011 to develop a first version of an atlas of background concentrations of various pollutants for Norway and thereby to update the previously used estimates of background concentrations based on the VLUFT tool. The report summarizing the work carried out in 2011 (Schneider et al., 2011) describes in detail the methodology used for this purpose. In addition, it lists several simplifications and error sources of the present methodology and suggests possible solutions. One of the recommendations in this report was to examine other auxiliary datasets such as satellite and model output for their potential to providing additional information to the methodology.

In 2012, additional funding was provided for evaluating additional auxiliary datasets regarding their potential of improving the spatial and temporal characterization of the background concentration estimates. This follow-up project focused on three major objectives:

1. To evaluate the potential of satellite-derived NO₂data for improving station- based NO₂mapping in Norway

2. To evaluate the feasibility of using data from a high-resolution atmospheric model for better spatial and temporal characterization of background concentrations

3. To make the results available through an online web mapping system to allow easy access to the data and to provide basic visualization of the results

The work carried out in this project in order to achieve these objectives is to a large extent based on the methodology developed in 2011. Therefore, a brief summary of the existing methodology is provided in the following. More details including a comprehensive description of the used data sets can be found in the previous report provided by Schneider et al. (2011).

The estimation of Norwegian background concentrations for NO₂, O₃, PM₁₀, and PM_2.5as carried out in 2011 is based on two main components. The first component consists of maps of the average annual background concentration for recent years that are derived from station observations in conjunction with spatially distributed auxiliary data using geostatistical techniques. However, since most of the pollutants considered here vary significantly with time, maps of annual averages alone are not sufficient. The second component of the methodology is therefore based on a quantitative description of the average long-term temporal behavior of the observations at each station (Schneider et al., 2011).

A combination of the two components was then accomplished within the framework of this project by averaging several years of hourly measurements on an annual as well as on a daily basis. The resulting time series for a typical year and a typical day were further smoothed to ensure that the observations are representative of cyclical temporal patterns and do not just reflect short-term variability. The representative annual and daily time series are subsequently converted from absolute concentrations given inµg m^-3to anomalies from the long-term mean at the station given in percent.

(12)

This ensures the applicability of the temporal information for neighboring areas with differing mean annual background concentrations.

Due to the often short time series available at each station and the associated small sample size, random noise which is not representative of the overall long-term temporal variability is abundant in the time series and needs to be removed before using the relative anomalies for estimating concentrations at other locations. Such a task can for example be performed by using a moving average filter. However, for practical purposes this smoothing was performed here in the operational application by applying a two-dimensional low-pass filter on an hour-by-hour anomaly matrix for an average year. This results in a simultaneous smoothing of both the annual and daily average time series. An example is shown in Figure 1. It should be noted that the application of the filter was performed while the matrix was augmented by itself on all four sides in order to avoid erroneous edge effects caused by the filter.

The smoothed relative anomalies can then be applied to neighboring locations with different absolute annual mean concentrations, and as such the average concentration can be estimated for a certain location given a certain day of the year and a time of day.

Figure 2 shows a comparison of the information content provided by the updated background concentration as opposed to the previously used 1993 VLUFT data set.

Compared to the previously used VLUFT dataset, the method presented here has clear advantages in that it provides a significantly higher information density in both the spatial as well as the temporal dimension. The method provides quantitatively reasonable estimates of background concentrations, although the uncertainty at the hourly level is quite high. The main source of uncertainty is the low number of suitable background stations located in Norway. A major advantage of the technique is further that it can be easily updated with new data. Figure 3 shows an overview of typical annual average concentrations of NO₂, O₃, PM₁₀, and PM_2.5, as generated for the previous 2011 study.

While the methodology devised for the previous project provided reasonable estimates of background concentrations in Norway, both the spatial and temporal components are associated with significant uncertainties. Schneider et al. (2011) list several potential improvements for reducing the errors, among others they recommended the evaluation of auxiliary datasets such as satellite data and high-resolution model output. These datasets were evaluated here and the methodology and results are described in the following sections.

(13)

JanFebMarAprMayJunJulAugSepOctNovDec

510

15

20

Conc. [µg m ] −3

0

20

40

60 JanFebMarAprMayJunJulAugSepOctNovDec

510

15

20

N (samples)

0

5

510

15

20

Anomaly [%]

−100

−50

0

50

510

15

20

Smoothed Anomaly [%]

−100

−50

0

50

100

a b c d

Time of day [hours] Figure1NOatstationNO0075ABarnehagen:Annualmatricesofhourlyaveragescomputedoverentireavailabletimeseries,shownasa)Observations,b)2 numberofyearswithavailabledata,c)theanomalycomputedfromthelong-termmean,andd)theanomalyfromthelong-termmeansmoothed usingalow-passfilter.

(14)

Figure 2 Comparison of the information content about background concentrations obtained from the previous method and the method described in this report, shown for the example of NO₂. Panel a) shows 1993 VLUFT data for rural areas for the medium-level class, panel b) shows the annual mean background concentrations for 2008 derived using the method presented here, panel c) shows an example of temporal information available from VLUFT, here for Akershus county, and panel d) shows the temporal concentration information at Kjeller in Akershus country for a typical year as derived by the method presented here. Note that the values from VLUFT given in panel a) are

”episodic high hourly concentrations“ and are thus not directly comparable to the annual mean values shown in panel b). (From Schneider et al. (2011))

(15)

Figure 3 Maps of 2007 (NO2) and 2008 annual means of NO2, O3, PM10, and PM2.5

as they were computed in the 2011 study. The spatial resolution of the grid is approximately 10 km×10 km. The maps are partly based on data provided by the European Topic Centre on Air and Climate Change implementing a methodology described in Horálek et al. (2010). (From Schneider et al.

(2011))

(16)

(17)

2 Data

A wide variety of data sources were used for this project. This includes hourly station data, satellite data, output from atmospheric chemistry models, and processed data from previous projects on mapping European air quality.

2.1 Station data

Raw data from air quality stations was used for both spatial mapping using residual kriging as well as for temporal decomposition of the time series. All station data was obtained from theEuropean Air quality dataBase, AirBase (http://acm.eionet.

europa.eu/databases/airbase/). However, different datasets were acquired for each component. For the geostatistical analysis, annual mean concentrations were acquired for all European background stations in order to achieve a large enough sample size for variogram modeling and regression analysis (see Figure 4). For the temporal characterization, only data for Norwegian stations were acquired for all four species, however this was done for the entire available record and at an hourly temporal resolution.

Figure 4 Map showing the 2009 average NO₂concentration measured at all Airbase background stations. (from Schneider et al. (2011))

(18)

Table 1 lists all background air quality stations located in Norway for which data was retrieved for the temporal component from the AirBase database. Traffic and industrial stations were not used because of their limited spatial representativeness. Therefore, only background stations (urban, suburban, and rural) were considered. The geo- graphical context is shown in Figure 5 which shows the location of all background air quality stations in Norway with suitably long time series for each component.

In addition, Table 2 gives an overview of station type and the components measured at each station with suitably long time series, as well as the respective long-term means for each component. Note that only a small number of stations provides suitable time series for NO₂and only one station provides data for PM_2.5. Swedish and Finnish stations were not used here for the temporal characterization but could provide valuable additional information in future work.

2.2 SatelliteNO₂ data

Operational satellite remote sensing of NO₂has been carried out since 1995 when the Global Ozone Monitoring Experiment (GOME) (Burrows et al., 1999; Richter and Bur- rows, 2002) was first launched. Beginning in 2002, the observations were continued by the SCIAMACHY (SCanning Imaging Absorption spectroMeter for Atmospheric CartograpHY) sensor onboard of Envisat (Bovensmann et al., 1999; Gottwald et al., 2006), and subsequently complemented in 2004 by the Ozone Monitoring Instrument (OMI) (Levelt et al., 2006) as well as the Global Ozone Monitoring Experiment-2 (GOME-2) instrument in 2006 (Munro et al., 2006).

Two satellite datasets of NO₂were evaluated here for their potential to be used within the context of European-scale air quality mapping. The first one was a monthly- averaged dataset acquired by the SCIAMACHY instrument and processed by the

Table 1 Overview of Norwegian background air quality stations that were used for temporal characterization. All station data was acquired from AirBase. Note that not all stations provide data for all air quality indicators and that stations not listed here were not considered due to short time series or other reasons.

(from Schneider et al. (2011))

Station ID Station Name City Lat.[deg] Long.[deg] Elevation[m]

NO0075A Barnehagen LILLEHAMMER 61.121 10.467 210

NO0001R Birkenes 58.383 8.250 190

NO0081A Bærum 59.952 9.645 80

NO0070A Grimmerhaugen AALESUND 62.472 6.166 21

NO0077A Gruben MO I RANA 66.310 14.194 10

NO0062A Haukenes 59.200 9.400 25

NO0056R Hurdal 60.367 11.067 300

NO0045R Jeløya 59.433 10.600 5

NO0055R Karasjok 69.467 25.217 333

NO0039R Kårvatn 62.783 8.883 210

NO0016A Nedre Storgate DRAMMEN 59.746 10.207 20

NO0041R Osen 61.250 11.783 440

NO0043R Prestebakke 59.000 11.533 160

NO0015A Rådhuset BERGEN 60.395 5.327 5

NO0052R Sandve 59.200 5.200 40

NO0072A Skøyen OSLO 59.920 10.733 10

NO0073A Sofienbergparken OSLO 59.356 10.766 25

NO0063A Stener Heyerdahl KRISTIANSAND 58.090 7.586 12

NO0015R Tustervatn 65.833 13.917 439

NO0065A Våland STAVANGER 58.961 5.731 33

NO0080A Øyekast 59.133 9.645 40

(19)

Figure 5 Location of the Norwegian background air quality stations whose data was used in this project for purposes of spatial mapping and temporal decomposition for a) NO₂, b) O₃, c) PM₁₀, and d) PM_2.5. The station type is indicated in the label as (u) for urban, (s) for suburban, and (r) for rural. Note that only stations with sufficiently long time series are shown.

Tropospheric Emission Monitoring Internet Service (TEMIS), which provides a comprehensive data archive at the websitetemis.nl. The retrieval algorithm used for the NO₂product investigated here is based on the methodology developed by Boersma et al. (2011) and is described in more detail in Section 3.1. Monthly global NO₂data for the entire lifetime of the SCIAMACHY instrument was available but only 2009 data has been used for comparison purposes here.

The second dataset tested here was acquired by the OMI instrument which flies onboard of NASA’s Aura platform. The specific dataset used was an experimental high-resolution product based on the OMNO2e Level 3 product. This dataset is produced at NASA Goddard based on the retrieval algorithm described by Bucsela et al. (2006) and Bucsela (2012). Global NO₂data with a daily sampling rate for the entire year of 2009 was available for this dataset.

Detailed information about the respective retrieval algorithms of the two satellite- based NO₂products are given in Section 3.1.

(20)

Table 2 Overview of station type and components measured at each station as well as their respective long-term mean. All means are given in units ofµg m^-3. When no annual mean is indicated the data either did not have sufficiently long time series for computing annual and daily means or the component was not measured at that station. The column CHIMERE indicates whether the station is located within the extent of the CHIMERE model output and thus is suitable for model comparisons. (from Schneider et al. (2011))

Station ID Station Name Type CHIMERE NO2 O3 PM10 PM2.5

NO0075A Barnehagen urban yes 19.2 - 19.0 8.8

NO0001R Birkenes rural no - 55.2 - -

NO0081A Bærum urban yes - 39.0 - -

NO0070A Grimmerhaugen urban no - - 13.1 -

NO0077A Gruben suburban no - - 17.4 -

NO0062A Haukenes suburban yes 5.6 54.8 - -

NO0056R Hurdal rural yes - 54.6 - -

NO0045R Jeløya rural yes - 56.1 - -

NO0055R Karasjok rural no - 65.7 - -

NO0039R Kårvatn rural no - 58.6 - -

NO0016A Nedre Storgate urban yes - - 19.9 -

NO0041R Osen rural yes - 55.8 - -

NO0043R Prestebakke rural yes - 58.5 - -

NO0015A Rådhuset urban yes 34.7 - 17.9 -

NO0052R Sandve rural yes - 66.2 - -

NO0072A Skøyen urban yes - - 21.8 -

NO0073A Sofienbergparken urban yes - - 22.0 -

NO0063A Stener Heyerdahl urban yes - - 22.1 -

NO0015R Tustervatn rural no - 70.0 - -

NO0065A Våland urban yes 16.7 - 15.8 -

NO0080A Øyekast urban yes 14.5 - 17.1 -

2.3 ETC/ACM data

As previously described in Schneider et al. (2011), existing data sets generated by the European Topic Centre on Air Pollution and Climate Change Mitigation (ETC/ACM) were used for the mapping component, whenever possible. The methodology un- derlying the mapping procedure has been refined over many years and the datasets have been extensively validated (Horálek et al., 2007, 2010; Denby et al., 2011).

Such data was available for NO₂, PM₁₀, and PM_2.5, however not for O₃. The annual average map for O₃over Norway was produced at NILU from raw datasets using a similar methodology.

The mapping methodology used by the ETC/ACM is described in detail in various reports, such as Horálek et al. (2007), Horálek et al. (2010), and Denby et al. (2011), and therefore will only be summarized here briefly. The approach uses a combination of a linear regression model which is then followed by the kriging of the resulting residuals, a process also known as residual kriging (Goovaerts, 1997). Separate maps are created for urban and rural areas which are later combined using specific merging rules based on population density. For each species and mapping type, a varying number of spatially exhaustive auxiliary variables are used which guide the interpolation process in areas of low station density. The type and number of auxiliary variables used within the mapping procedure is dependent on their respective impact to an improved fit of the regression model. For example, the interpolation of PM₁₀ in rural areas used output from the EMEP model, a digital elevation model for information on altitude, data on wind speed, and data on solar radiation. On the other hand, for PM₁₀ mapping in urban areas the used auxiliary variables consisted solely of the output from the EMEP model. For more detail on the auxiliary variables

(21)

used for the mapping of NO₂, PM₁₀, and PM_2.5see the reports provided by Horálek et al. (2007), Horálek et al. (2010) and Denby et al. (2011).

Once the multiple linear regression against the appropriate auxiliary variables is accomplished, residuals are acquired at each location where station data is available.

These residuals are subsequently interpolated using ordinary kriging (Cressie, 1993;

Goovaerts, 1997; Wackernagel, 2003). This interpolation process is based on variogram analysis, according to which the spatial autocorrelation of the data is fitted using a (often spherical) variogram model. Kriging weights are obtained as a result of this process and the optimal prediction of residual concentration is made at each 10 km×10 km grid cell. Subsequently, a final map of estimated concentrations is obtained by adding the gridded result from the linear regression and from the kriging of the residuals.

In addition to the linear regression and ordinary kriging techniques resulting in estimated concentration maps for rural and urban areas, the ETC/ACM methodology further uses a fairly sophisticated merging procedure for combining the separately interpolated maps of urban and rural areas. The technique is based on the population density for each grid cell and assign the interpolated value from the rural map if the population density is less than a given thresholdα1and assigns the interpolated urban value for all cells exceeding a population density ofα2. In case the population density is greater than α1 but less thanα2, a joint rural/urban value is computed using a weighting function and assigned to the respective grid cell. Once all the grid cells are assigned their appropriate concentration values based on their respective population density, a final concentration map of the parameter in question is obtained.

2.4 CHIMERE Model Data

A one-year high-resolution run of the three-dimensional chemistry-transport model CHIMERE (Schmidt et al., 2001; Vautard et al., 2001; Vautard, 2003; Bessagnet et al., 2004) was used for this part of the study.

Originally developed as an extension of a regional-scale model developed for the Paris area (Vautard et al., 2001), CHIMERE is a multi-scale chemical transport model which is primarily designed to generate daily forecasts of ozone, aerosols and other pollutants, as well as for producing long-term simulations for the purpose of emission control scenarios.

Europe-wide CHIMERE output for the entire year 2009 with an hourly sampling rate and a spatial resolution of 0.0625°×0.125°(approximately 7 km×7 km) was available for this study. Unfortunately, the northernmost extent of the model domain was 61.8°N, which does not include all of Norway. However, most of southern Norway, which includes Norway’s two largest urban areas Oslo and Bergen is included in the model domain. It was therefore decided to go ahead and test the methodology based on this region even though it was not possible to integrate the dataset in the operational mapping procedure for all of Norway.

(22)

(23)

3 Methodology

3.1 Satellite Retrieval Methodology

Two satellite NO₂products were further investigated for their potential use within this study. The first product tested was acquired by the SCIAMACHY instrument onboard of the Envisat platform. The product used is based on the TEMIS retrieval algorithm (Boersma et al., 2011).

In short, the TEMIS NO₂ retrieval is based on three steps: The first step of the algorithm consists of a Differential Optical Absorption Spectroscopy (DOAS) retrieval of the total slant column of NO₂ from the measured spectrum, where absorption cross sections of NO₂, ozone, H₂O as well as a synthetic ring spectrum are taken into account, and a fifth order polynomial is included in the fit to account for scattering effects. The second step consists of the separation of the stratospheric and tropospheric NO₂ contributions to the total NO₂column, where the stratospheric NO₂column is estimated by assimilating total slant columns in the TM4 chemistry transport model (Dentener et al., 2003; Boersma et al., 2007). The third and final step of the retrieval is the conversion of the tropospheric NO₂slant columns into vertical columns using a calculated Air-Mass Factor (AMF). Further details on the specific retrieval methodology can be found in Boersma et al. (2004), Boersma et al. (2007), and Boersma et al. (2011), as well as on the TEMIS website (www.temis.nl^).

Solely data reprocessed with version 2.0 of the retrieval algorithm was used. Im- provements in version 2.0 over previous versions of the retrieval algorithm include an updated albedo database, a modified calculation of the air mass factor, a correction of the surface height calculation, a correction of the weekly cycle in NO_xemissions, as well as an increased number of NO_xtracers in the applied chemical transport model (Boersma et al., 2011). The NO₂dataset used here only considered cloud radiance fractions of less than 50%. It was also resampled from the original SCIAMACHY spatial resolution to a 0.25 degree×0.25 degree grid.

Although the TEMIS-based NO₂dataset used in this study is based to some extent on data assimilation using the TM4 model (Dentener et al., 2003; Boersma et al., 2007), it is almost independent of the used emission inventory due to the retrieval set-up. The data assimilation results are mainly used to provide the stratospheric NO₂ column in the second step. This stratospheric column is virtually independent of the used emission database. For the calculation of the AMF in the third step knowledge of the profile shape of the vertical NO₂distribution is needed. This profile shape is also taken from the data assimilation. However, the profile shape is independent of the emissions, since the data assimilation is scaling the NO₂column with conservation of the shape. In conclusion, the NO₂data are considered as retrieval results independent of emission data.

The second satellite NO₂product tested here was acquired by the Ozone Mapping Instrument onboard the Aura satellite. The OMI product used within the framework of this study is based on a retrieval algorithm developed at NASA (Chance, 2002).

The original, version 1 retrieval algorithm is described in Bucsela et al. (2006). The new version 3.0 retrieval algorithm is greatly improved over the previous versions (Bucsela, 2012).

The retrieval algorithm for the OMI NO₂product consists of a total number of four major steps: A Differential Optical Absorption Spectroscopy, a calculation of the

(24)

air mass factor, destriping, and a troposphere-stratosphere separation. The DOAS analysis first divides earthshine radiances by the reference solar irradiance spectrum.

The normalized spectra are then fitted to trace gas spectra observed in the laboratory using a reference Ring spectrum and a polynomial function. The DOAS fitting is applied in the spectral range of 405 nm to 465 nm. In a next step, the air mass factor is calculated using scattering weights and a monthly mean climatology of NO₂ profile shapes, which were derived from a chemical transport model. The AMF is subsequently computed using the cloud radiance fraction f as

AM F = (1−f)·AM F_{cl ear}+f ·AM F_{cl oud} (1)

whereAM F_{cl ear} andAM F_{cl oud}are the model-derived air mass factors for clear and cloudy conditions, respectively. Following the AMF calculation, the NO₂slant column densities observed by OMI are then “destriped” in order to correct for an instrument artifact. Finally, as a fourth step, a troposphere-stratosphere separation is performed using ana prioriestimate of the tropospheric contribution based on a monthly model climatology.

More information about the OMI NASA retrieval algorithm can be found in Bucsela et al. (2006), Bucsela (2012), and OMI Team (2012). The OMINO2 product (Chance, 2002) is estimated to have a fitting error in the slant column of approximately 0.3 - 1

×10¹⁵ molecules cm^-2 (OMI Team, 2012).

3.2 Geostatistical framework

The European background maps are created using a geostatistical technique, namely residual kriging with auxiliary variables. Kriging is an interpolation technique that makes use of a model of spatial autocorrelation (usually in the form of a variogram model) to infer optimal estimates of a variable at a given set of locations (Isaaks and Srivastava, 1989; Cressie, 1993; Goovaerts, 1997; Wackernagel, 2003).

The mapping procedure applied in this study is based on the previous work by Horálek et al. (2007), Horálek et al. (2010), and Denby et al. (2011) and involves a linear regression analysis against an auxiliary variable in conjunction with kriging of the residuals. It should be noted that the cited work incorporates a procedure for separately mapping urban and rural areas and then combining the interpolated maps using a merging technique. This part of the algorithm was not implemented in the mapping procedure for this project.

The concentration ˆZ(s₀)is mapped at a given locations₀using the model

Zˆ(s₀) =c+a₁X₁(s₀) +a₂X₂(s₀) +. . .+a_nX_n(s₀) +η(s₀) (2) wherec,a₁,a₂. . .a_nare parameters of the multiple linear regression andX₁(s₀). . .X_n(s₀) are the values of the auxiliary variables used at locations₀. Finally,η(s₀)represents the results of the ordinary kriging of the residuals at locations₀. While equation 2 provides a general methodology for incorporating multiple auxiliary variables, only single auxiliary variables were tested here in order to evaluate the impact of each auxiliary variable individually (with one exception mentioned later on). The first step in the process was therefore to establish a linear relationship between the annual average NO₂concentration at each station and the respective auxiliary variable at each station. This task was performed throughout all background stations in Europe available within AirBase (with exception of those stations used for validation) in order to obtain a representative relationship.

(25)

0 2 4 6 8 10 12 14 70

75 80 85 90 95 100 105 110 115 120 125

Lag Distance h [deg]

γ(h)

Figure 6 Example of an empirical semivariogram ˆγ(h)and its model, describing the autocorrelation of a European NO₂station dataset. The model in this case is a combination of a nugget effect of 74.7 and a spherical model with sill 40.6 and a range of 14 degrees.

Kriging makes use of a model describing the spatial autocorrelation. Most often, the semivariogramγ(h)at a certain lag distancehis used to describe this. Different types of models are then fitted to the empirical semivariogram, with a spherical and Gaussian models probably being the most common. Figure 6 shows an example of the empirical semivariogram and the fitted spherical model used for residual kriging of NO₂over Europe.

For kriging of residuals, a model was fitted to the empirical semivariogram of the residuals with a combination of a nugget effect model and a spherical or Gaussian model of range a₀degrees and sillc₀µg m^-3 such that the semivariance ˆγat laghis given as either

γ(ˆ h) =





 c₀·

3 2

h a₀ −¹₂

h a₀

3

for h≤a₀

c₀ for h>a₀

(3)

for the spherical model or

γ(ˆ h) =c_o

1−e x p

−h² a²₀

(4) for the Gaussian model.

The fitted semivariogram model is then used in the kriging process to determine appropriate weighting factors for each data point. More detailed information about the kriging process can be found in the literature, e.g. in Isaaks and Srivastava (1989), Cressie (1993), or Goovaerts (1997). The kriged residuals are then added to the results from the multiple linear regression as indicated in Equation 2 and through this process the final results are obtained.

Figure 7 illustrates the basic workflow using a schematic of the methodology for the case of using satellite data as an auxiliary variable.

(26)

Station data 0.1 deg OMI NO2 data

Establish correlation

Semivariogram modeling

Kriging of residuals Final NO2 map

Combineregression result and krigedresiduals Compute mapusing only regression Figure7Schematicofthegeneralmethodology,hereillustratedontheexampleofusingOMINO2satellitedataasanauxiliarydatasetforkrigingstationobservationsofNO2.

(27)

4 Results and Discussion

In the following sections the results of the three main tasks of this project are presented and the implications discussed. First, the impact of integrating satellite data of NO₂ in the mapping procedure is described. Subsequently, the potential of high-resolution output of a chemical transport model is evaluated. Finally, a web-based mapping tool for visualizing and accessing the spatial and temporal information in the dataset of background concentrations over Norway is presented.

4.1 NO₂ mapping in Norway using satellite data

Based on a growing importance of spaceborne data for air quality related applications it is highly desirable to study the impact of satellite data on currently existing air quality mapping techniques. In this section, the potential of using satellite-based NO₂ data as an auxiliary variable for mapping air quality at the European and Norwegian scale using geostatistical techniques is investigated. Two satellite products were compared and one was selected for further use within the actual mapping procedure.

4.1.1 Choosing a suitable satellite product

As a first objective of this task, a satellite product suitable for use within the mapping procedure had to be found. For this purpose, two satellite-based NO₂products were chosen for further investigation: The SCIAMACHY product based on the algorithm by TEMIS (Boersma et al., 2011), and the OMNO2e product (Bucsela et al., 2006;

Bucsela, 2012). Figures 8 and 9 show the 2009 annual mean tropospheric NO₂ columns derived from the OMI and SCIAMACHY products, respectively, each using different retrieval algorithms. Note that the color scale on both figures is identical, so both qualitative and quantitative comparisons can be carried out.

Overall, the spatial patterns shown by the two products agree quite well. All the major regions of generally high NO₂concentrations, such as the region of Belgium and the Netherlands, southern and eastern England, as well as the Po valley region in Northern Italy, are captured adequately by both products. Furthermore, individual NO₂ hotspots over more isolated cities such as Moscow, Madrid, and Istanbul are easily identifiable from both data products. The map produced from OMI data appears to be slightly smoother whereas the SCIAMACHY-based maps shows a bit more “noise”. This is due to the fact that the OMI-based annual mean map was computed by averaging over daily images, whereas the SCIAMACHY-based annual mean was calculated from monthly average datasets, which in turn were derived from daily data.

In terms of actual NO₂concentrations, it is obvious from the two figures that SCIA- MACHY overall measures significantly higher columns in the polluted areas than OMI.

Figure 9 clearly shows this effect as significantly larger areas exceeding 10×10¹⁵ molecules cm^-2 as compared to 8. This effect is particularly obvious in the Po valley region in Northern Italy, for which the OMI annual mean only shows very few grid cells exceeding 10×10¹⁵molecules cm^-2, whereas most of the region of Northern Italy exceeds this value in the SCIAMACHY-based map.

The reason for this behavior can be found in the combination of the strong diurnal cycle of NO₂ in heavily polluted areas and the different overpass times of the two

(28)

Figure 8 Annual mean NO₂concentration for the year 2009 derived from the OMNO2e daily 0.25°×0.25° product. Note that the overpass time of the Aura platform on which OMI is flying, is at approximately 13:30 local time.

instruments. While the Envisat satellite, on which the SCIAMACHY instrument is mounted, has a local overpass time at the equator of around 10:00 local solar time (LST), and thus samples the tail end of the morning rush hour, the OMI instrument on the Aura platform has a local overpass time at the equator of approximately 13:45 LST and as such samples the atmosphere in the middle between the morning and evening rush hours. As such, its observations of tropospheric NO₂columns are expected to be lower than those obtained from SCIAMACHY.

In order to explore the quantitative difference between the two products in more detail and with a particular focus on spatial patterns, a difference image between the products from the two instruments was produced. The difference in NO₂column∆C given in×10¹⁵molecules cm^-2 was calculated as

∆C=CSC I AM AC H Y −C_{OM I} (5)

whereCSC I AM AC H Y andC_{OM I} are the annual mean NO₂column for SCIAMACHY and OMI, respectively. Based on this equation, positive values in the difference image indicate that the SCIAMACHY retrieval is higher than the OMI retrieval, and negative values indicate the opposite.

Figure 10 shows the resulting difference map. As expected, the highest absolute differences can be found over the most highly polluted areas. In Northern Italy, which

(29)

Figure 9 Annual mean NO₂concentration for the year 2009 derived from the SCIA- MACHY/TEMIS monthly 0.25°×0.25° product. Note that the overpass time of the Envisat platform on which SCIAMACHY is flying, is at approximately 10:00 local time.

exhibits the largest area of substantial differences, the values easily reach and exceed 5×10¹⁵molecules cm^-2. Several regions in Germany, Belgium, the Netherlands, and the United Kingdom also reach such high values, albeit only in areas of considerably smaller spatial extent. SCIAMACHY generally shows higher tropospheric columns by approximately 1×10¹⁵ molecules cm^-2 on average over large areas of Eastern Europe, particularly in the Ukraine.

In areas of generally low tropospheric NO₂concentrations such as over the oceans, Scandinavia, and Africa, OMI exhibits slightly higher values by approximately 0.5

× 10¹⁵ molecules cm^-2. However, this magnitude is easily within the error range specified for the products and thus probably is not of too much significance.

It should be noted that, while such an inter-comparison between two satellite products is not a substitute for validation with in situ data as it can not provide an absolute error estimate, it can provide valuable information on spatial patterns in differences.

Despite differences in absolute values, it is critical to point out that the spatial patterns indicated by both instruments are very consistent. This is important because when using such satellite-based maps as an auxiliary datasets for supporting kriging of station data, it is primarily the spatial patterns that have an impact on the results, whereas the absolute values are based on the station data.

(30)

Figure 10 Difference image of the mean annual NO₂ column retrieved from SCIA- MACHY and OMI. The difference is calculated based on Equation 5. Note that both satellite instruments have significantly different overpass times (10:00 vs. 13:30 local time), which together with the diurnal cycle of NO₂ explain the majority of the inter-sensor biases.

While for the previous figures and analysis the 0.25°×0.25° resolution OMI product was used to provide as much consistency as possible with the SCIAMACHY product, a high-resolution 0.1°×0.1° OMNO2e product exists for the OMI instrument. Given the similarity in spatial scale between the 0.1°×0.1° OMNO2e product and the 10 km spatial resolution at which air quality is being mapped operationally in Europe by the ETC/ACM, this product is a natural choice for this study. Figure 12 shows a direct comparison of the two OMI products. The high resolution product clearly can resolve more detail and provides higher values in some hotspots which do not appear in the 0.25°×0.25° resolution product due to spatial averaging.

Based on these results and further based on the fact that a 0.1°×0.1° product was available from OMI while only 0.25° ×0.25° resolution was available from SCIA- MACHY, it was decided to use the OMNO2e product for the remainder of this study.

The relatively high resolution of the OMNO2e product allows for mapping at the 10 km grid cell level for all of Europe. It should further be noted that, in contrast to for example SCIAMACHY data, OMI observations are available at present and further will be continued at an even higher spatial resolution in 2014 with the launch of the TROPOMI instrument onboard the Sentinel-5 precursor platform.

(31)

Figure 11 The 0.1°×0.1° resolution OMNO2e product over Europe. Shown here is the 2009 annual mean tropospheric NO₂concentration. Significantly more detail is visible than in the standard resolution product (see also Figure 12). See also the same figure with adjusted extent and color scale showing Norway only (Figure 13).

(32)

Figure 12 Comparison of the 0.25°×0.25° resolution OMNO2e product (top) with the 0.1 degree resolution OMNO2e product (bottom), shown for the Po valley region in Northern Italy. The higher resolution product clearly shows details not visible in the image of the 0.25°×0.25° resolution product. The figures show the annual mean tropospheric NO₂column in 2009.

(33)

Figure 13 Annual average tropospheric NO2column for the year 2009 over Norway.

Derived from the 0.1 degree high-resolution OMNO2e product. This figure shows the same data as displayed in Figure 11 however it uses a modified color scale in order to highlight spatial patterns of NO2in Norway.

4.1.2 KrigingNO₂in Norway using Airbase and OMI satellite data

Based on the results reported on in the previous section and in order to provide an indication as to what extent satellite data of NO₂can help to improve European-scale mapping of air quality, OMI-based tropospheric column NO₂data was subsequently used in the next step to complement the station measurements from Airbase (see Figure 4) as an auxiliary dataset. As described in detail in the methodology section, this was accomplished by establishing a correlation between the station-based NO₂ means and the mean satellite-based tropospheric columns as observed over each station.

(34)

Figure 14 Scatterplot of Airbase-derived annual mean 2009 station NO₂concentration against the 2009 annual mean tropospheric NO₂columns derived from the OMNO2e high-resolution product.

Figure 13 shows the 2009 average tropospheric NO₂ column over Norway based on the OMNO2e 0.1°×0.1° resolution satellite product. This map is similar to the one shown in Figure 11 but it has a modified color scale in order to better highlight the spatial patterns within the overall low values of the tropospheric NO₂column in Norway.

Figure 14 shows a scatter plot indicating the correlation between the 2009 annual average NO₂concentration at all background Airbase stations and the 2009 annual average tropospheric column extracted at each station location from the high-resolution annual average OMI dataset. A linear model was fitted to this dataset as

C_{OM I}=1.89+0.12×C_{S t} (6)

whereC_{OM I} is the tropospheric column observed by the OMI instrument andC_{S t} is the annual mean NO₂concentration observed at each Airbase station. The R²value of the model was found to be close to 0.3.

At first glance this correlation might appear to be quite weak, however it needs to be considered that this analysis compares two parameters which have very different spatial and temporal scales. While the station observations provide an annual mean NO₂value at the ground level computed from hourly values and which is representative of only a very small area, the satellite instrument provides the total number of NO₂ molecules at 14:00 local time, averaged not only over a 100 km²area but also integrated over the entire troposphere. Given these fundamental differences in spatial and temporal scales, the correlation seen in Figure 14 is quite reasonable.

The residuals resulting from the fitted linear regression are shown in Figure 16 for all of Europe. As mentioned before, while the goal of this project was to map background concentrations in Norway, the geostatistical analysis had to be performed

(35)

0 2 4 6 8 10 60

65 70 75 80 85 90

lag distance h (deg)

γ(h)

Figure 15 Empirical and modeled semivariogram of the residuals (shown in Figure 16). The model is a combination of a nugget effect of 64.3 and a Gaussian model with range 5.9 and a sill of 18.9.

at the European level in order to obtain enough sampling points for deriving a representative semivariogram model. In addition, Figure 17 shows the residuals obtained for air quality stations located in Norway only for reference.

Overall, the spatial patterns in the residuals indicate negative values throughout the northern part of Central Europe, i.e. in Germany, the Netherlands, Belgium, and the western part of Poland. Positive residuals can be found throughout most of the rest of Europe, however the highest density of positive residuals occurs in Northern Italy and the east of France. In Norway, the situation is mixed, with stations in Oslo, Kristiansand, and Bergen, and Lillehammer showing positive residuals while the rest of the stations has negative residuals. In particular the entire northern half of the country exhibits negative residuals.

The residuals were then subsequently plotted as an empirical semivariogram, which was then in turn modeled using a combined nugget effect of 64.3 and a Gaussian model with range 5.9 degrees and a sill of 18.9 (see Figure 15). The semivariogram model was then used to krige the residuals from the previously discussed linear regression over the entire study domain. This domain ranged from 20° N to 73° N and from 20° W to 40° E. A spatial resolution of 0.1° was used for the final grid. Figure 18 shows the kriged residuals over all of Europe.

A final map of NO₂ was then generated by combining the regressed map with the map of the kriged residuals. The result of this effort is shown in Figure 19. As would be expected, the spatial patterns of NO₂in Norway contain elements of both the OMI satellite dataset and the kriged residuals of the station observations.

When qualitatively comparing the spatial patterns found in the original map of the satellite measurements of NO₂tropospheric column (Figure 13) with the final map of NO₂found in Figure 19, it can be observed that the southwestern part of the country now has slightly higher values than before due to the impact of the positive residuals (which were primarily caused by a strong positive value at the Bergen station).

(36)

Figure 16 Map showing the residuals from the model fitted between the average 2009 NO₂at at all Airbase background stations and the mean tropospheric NO₂ column provided by the high-resolution OMNO2e product.

A very preliminary validation of the data was carried out at the European level. Cross- validation was used to evaluate the quality of the results. Based on this validation technique, the original Airbase dataset was randomly split up in two parts. The first part, encompassing 90% of the stations, was used within the kriging procedure. The second part, consisting of approximately 10% of the Airbase stations was used solely for validation purposes. This procedure ensures that the stations used for validation had absolutely no impact on the quality of the result as they were not used as part of the algorithm. This resulted in a total number of 198 randomly selected stations that were separated from the main Airbase dataset and only used for validation purposes.

The validation for this map resulted in an RMSE of 8.5µg m^-3 at the European level, which is lower than the mapping carried out using solely station data (RMSE=9.1 µg m^-3). As expected, this indicates that the satellite dataset provides additional valuable information on spatial patterns.

The results of this part of the study indicate that satellite data can be very useful as an auxiliary variable in mapping air quality at both the European and Norwegian scale. Using tropospheric column NO₂data acquired by the OMI instrument provided significantly better mapping results (both qualitatively as well as quantitatively) than geostatistical interpolation of station data alone.